From Layouts to Physics: A Novel Modeling Approach for Analog Layout Automation

Analog layout automation has long been a central goal in electrical design automation. However, learning-based analog layout automation remains difficult to scale. Existing approaches often start from schematic netlists and attempt to learn layout generation or layout-quality prediction, but this forward direction is intrinsically ambiguous: the same netlist can admit many valid layouts, while high-quality schematic-layout-QoR triples are scarce, expensive, and often proprietary. Moreover, circuit-level layout quality is difficult to define without a valid testbench and costly post-layout simulation, making large-scale supervision impractical. This paper proposes L2P, a layout-first, two-stage modeling paradigm for analog layout automation. Instead of treating generated layouts as unique golden labels for netlist-to-layout imitation, we use layouts themselves as deterministic sources of physical supervision. Under a fixed PDK and extraction setup, each layout can be mapped to a unique extracted implementation graph and a parasitic profile, revealing the physical consequences of placement and routing decisions. L2P first learns layout-to-physics relationships from abundant open-source layout and extraction data, and then uses a small amount of golden testbench-labeled data to align these physical quantities with circuit-level quality metrics. Across 12 downstream QoR tasks, L2P reduces simulated-label requirements by up to 96.8\% while matching training-from-scratch performance, reaching or surpassing the performance of conventional forward models trained on full datasets. L2P further supports bidirectional transfer of the same tasks across distinct design sets and PDKs.